teal-sea / zeta-labstate of record · compiled 14 Aug 2026 · revision 9ebdea0 · source

Library · hunts/frontier_math/START-HERE.md

START HERE — the honest state of the frontier_math hunt

1,034 words · 127 lines · source

Written 2026-08-13 as a handoff, because the working record grew to 29 markdown files and 8131 lines and buried the one thing that matters. Read this file. Ignore the rest unless it is cited from here.

Corrected 2026-08-14, because a page that exists to carry the one number that matters carried the wrong one. See the correction box in §2.


1. The objective

Improve the constant in "More than two thirds of the zeros of the Riemann zeta function lie on the critical line" (10 Aug 2026).

paper: 0.6725007037

2. Have we improved it? Not proved. A candidate.

reading of record: 0.6725106958 (+9.99e-6), a CANDIDATE

It is a candidate, not a result: one step of the chain is open, so the composite takes that grade. An earlier, larger candidate from this same programme (0.672529, later 0.6725124) was killed by this project's own controls, see PROOF-LEDGER.md line 24, "CLEAN KILL". That is the base rate to keep in mind.

Correction, 2026-08-14. The first version of this page named 0.6725087070 (+8.00e-6) as the reading of record, and said the higher figure "waits on a burden that has not landed". That was the ledger state at PROOF-LEDGER.md line 438, and it had already been superseded when this page was written. Burden (a), the chain re-run at the paper field, is delivered in the section beginning line 449, and line 462 records "Reading of record | MOVES to 0.6725106958"; lines 529, 574, 593, 609 and 660 hold it there. README.md, docs/27 and the public site carried the right number and this page did not. It is the same shape as coordinator defect #23, logged two screens further down in the same ledger: a superseded figure carried past its own retraction, by a session reading one row instead of the last row.

Four pieces of the chain are kernel-checked and none of them is in §3 below: the census floor (FloorCert.lean), the retention certificate's arithmetic (BandCert), the composition inequality (t3_composition_skeleton.lean) and the grid-incidence law (law_d_incidence.lean). Statements and obligations in docs/27 §2.

Distance to a clear win, honestly: one open research question (§4). The Lean gap this page listed beside it has closed since: the arm was built for the first time on 2026-08-13, it refuted the O9 leaf table on 7 of 9 chunks, and the repaired 699-cell table now decides at kernel grade on all 18. The research question still has no schedule.

3. What is actually solid

Grade hardened = measured many ways with controls that have power; not kernel-checked. Nothing below is kernel-checked.

  1. The k = 1 retention inequality, for every n, every shift, every depth in [0,1/2], with no separation hypothesis. Four independent instruments agree, plus an exact-rational certificate and a from-definitions reproduction (salvage_audit.py, 7/7). This is the real result of the session.
  2. ghat(z) = Phi2(-i z) — the BandCert and EForm3 arms compute one function. Residual exactly 0.0. Puts 7 of 12 Road A obligations on machinery that already compiles.
  3. The k-pair identity (residual 4.21e-17) and its three-term Gram form, two of whose three terms are free.
  4. int D(y,s) ds = -2 pi c2(0), exact and depth-free.

4. What is open

5. What to ignore

Most of the 29 markdown files are working record, not input. If you read five, read: this file, RETENTION-PROBLEM.md (the prover-facing statement), O9-SCOPING.md (the O9 route), O9-LEAF-REPAIR.md (what the kernel did to it, and the current cell counts), ACTIVE-CLAIMS.md (who else is editing this directory).

PROOF-LEDGER.md is ~2000 lines and is a defect record, not a plan. Five coordinator defects (#19-#23) were logged in one session, all the same shape: a quantity carried across contexts without being re-derived in the context it was used. That record is worth keeping and is not worth reading end to end.

6. The one next action

Superseded, and worth reading for why. This page originally said: get a container with elan + Lean 4 v4.33.0-rc2, build the 2-D O9 table and run lake build, because "one build converts a pile of staged work into a list of what actually fails". That happened on 2026-08-13 and the list was short and expensive: decide +kernel refuted the table on 7 of 9 chunks. The generator computed its transcendental leaves with Arb at 300 bits and a 4-ulp pad while Leaves.lean builds them from truncated series, so the model was narrower than the kernel and the correction cost roughly a factor of two in cells, 339 to 699. The prediction to keep is the general one in O9-LEAF-REPAIR.md §2: a generator with no round-trip against the thing it claims to mirror can only ever confirm itself.

The next action now is the denominator bridge in §4, and after it the Dam y s <= c*y^2 statement O9 exists to support.

k >= 2 is a separate, open-ended research effort and should not be sequenced behind that.

7. What went wrong in the session that produced this

Stated plainly so it is not repeated: the bottom line — we have not moved the constant — was never surfaced in one place. It was true throughout and buried under working record. A large share of the session went into correcting my own errors and re-sizing the same table three times. The markdown sprawl is self-inflicted. This file exists to replace it.